> Some years ago, a police officer casually told me, “You should get a dashcam, insurance fraud is common around here.” His offhand comment stuck with me, but life moved on.

This is very, very good advice. Dashcams are cheap, and easy to install. Treat yourself - spending $80 on a camera could save you thousands in lawsuits, insurance hikes, costs of a new car, etc.

To me it seems nearly as important as having a smoke detector in your house.

Lol, wow! The article says

>After filtering out invalid VINs, we narrowed the dataset to roughly ~15 million crashes. (We also removed all drug and alcohol related crashes, since it’s unlikely someone committing insurance fraud would be under the influence.)

Which I immediately found extremely implausible. Seems like the author is very much NOT familiar with the type of people who commit insurance fraud.

In the next paragraph the article links to the FBI website on the fraud ring in Connecticut.

https://www.fbi.gov/news/stories/staged-accident-ring

Which says

>After an autumn evening of drinking and using drugs in 2013, a group of friends got into an Audi A6 and drove to the remote Wilderness Road in Norwich, Connecticut. The car slid off the road, hitting a tree.

>Everyone in the car survived, but this seemingly typical crash was no accident.

>Despite their impairment, the driver and passengers had purposely planned the crash to collect the insurance money.

So yeah, don't exclude drug/alcohol related crashes. Also the author should read the first fucking sentence of the stuff they link. It's only six words in!

It’s very unfortunate that the worst people in society ruin insurance for everyone else - so that when you need to use it you are investigated as a criminal. Too many systems are being broken by a tiny minority of people with no skin in the game and no social consciousness.

Why would it be unlikely that the person committing fraud is under the influence? There's no reason to assume that.

It's an interesting article, but the title ('we found fraud!') doesn't quite match the article, which continually points out that there are suspicious patterns in the data but that they aren't ironclad evidence of fraud.

That said ... yeah those particular VINs do look seriously dodgy to me! If nothing else, then the drivers reporting 5 crashes in a 12 month window probably shouldn't be driving.

I'm kind of irritated that the article doesn't say anything about probabilities.

Given they have enough data, at some point it's perfectly reasonable to have cars with 5+ crashes per 12 months - just because of chance.

This is exactly why statistics was invented, damnit!

I don't see nearly enough complexity in this analysis to justify the claim of having found any insurance fraud.

Firstly, there's no account for correlation between the features identified. The article mentions VINs which have several single-vehicle accidents, for example, but someone who has one single-vehicle accident is probably more likely to have another. Switching coverage is another of those potentially-correlated features; if you claim and it bumps your premium, aren't you likely to shop around as a result?

Secondly, there's no attempt to account for the law of large numbers. It's incredibly unlikely that someone has three single vehicle accidents in a year, but because the probability of that is nonzero, we know that with enough vehicles on the road then someone is going to do it.

The article covers itself by acknowledging this, of course, but if you title your blog post "We Found Insurance Fraud in Our Crash Data" then you should actually do that.

The conclusion is a little too strong for my tastes. Changing coverage looks sketchy, but I’d bet people shop around after a crash and the ensuing rate hike.

Being involved in multiple accidents could suggest fraud but it could also just be bad driving, maybe someone very young or old.

It’d really help to know the base rates for many of these things: how often do people switch insurance, for example?

So many alternative explanations unmentioned:

1. Drink/drugs impair your judgment, making you more likely to commit fraud if you were on the fence, or the impulsive sort.

2. Drunk drivers get into multiple accidents, but some are able to keep evidence of impairment off the record (buddy on the force or the insurance co, etc.)

3. Drivers doing drugs that aren't detectable on the spot, and clear the body quickly, reducing evidence of drug use in crashes.

4. More people are likely to do drink/drugs at night, so if 1-3 are true, we should expect to see more crashes at night. Likewise for a lack of witnesses (fewer passengers or bystanders).

Most importantly, statistics are completely absent from the article. Does OP have any evidence these numbers wouldn't fit the correct random distribution? I know the point of fraud is to appear natural, but if done well, that means evidence has to be collected by other means, and this kind of analysis is only a weak signal.

New Title: "We Can't Prove Insurance Fraud with Our Data"

They just detected fraud suspicions and haven't found it yet. And this is not their business to detect fraud, that's up to investigators.

And it's not very wise to announce your metrics publicly. All fraudsters can read also.

I'm a little skeptical, it would be interesting to add odometer data, time of day, and distance from home. My driving is on low end of the collision risk spectrum because I don't drive much, mostly day time, and mostly routes that I've driven hundreds or thousands of times and where know the lines of sight and cadence of traffic. On the other hand I would expect more risk for someone who worked night shifts, did taxi or delivery type driving where they're often in unfamiliar areas, or just plain put in more miles. I would expect these effects to create a wide collision frequency spread even holding constant skill and good intentions.

> we never expected that a side by side look at Vehicle Identification Numbers and crash timelines would hint at possible insurance fraud.

Uh... I'm not surprised at all that auto insurance fraud, which involves vehicle collisions, might show up in vehicle collision data. This is sensationalist crap.

I expected this to be about data found in the memory dump of a crashed process. (And was surprised they would post about this at all, given privacy issues.)

As others have pointed out VIN as your primary level of detail is fairly flawed. You should also be filtering on vehicle registrations. A possible scenario here is someone owns their car, gets into an accident, repairs it or sells it off (insurance sell off). Someone else buys it gets difference insurance, and has an accident because its a new car to them. You would never know this vehicle story without integrating the registration history or at least buy and sell dates to know it has exchanged hands.

Yes Fraud is tricky, but a VIN does not equal a person (PII) committing the behavior, and these poor association attempts leave innocent people screwed by insurance companies. What makes this bad analysis annoying, is the constant caveats, if you know VIN associations don't prove fraud then maybe don't build some sort of risk scoring on it, and then try an sell it.

After reading the comments, I’d like to clarify a few things. This article isn’t meant to pitch a fraud scoring product. I recently discovered that several states publish VIN level data and, as a weekend project, decided to dig a bit deeper to see what I could learn. Given more time, I’d refine the analysis by normalizing against registration volumes, flagging outliers, and so on but for now I simply want to share an interesting glimpse at what ive found.